
Building STACKR: Giving Informal MSMEs a Financial Voice, Powered by AWS
STACKR transforms informal business activities such as paper bills and natural-language voice entries into verified financial records, business insights, and lender-ready financial documentation.
The Problem That Started It
Walk into any kirana store, corner shop, or local vendor stall across India, and you'll find a business that's very much alive — inventory moving, customers returning, credit extended to regulars on trust. What you usually won't find is a record of any of it. Sales live in a notebook, if they're written down at all. Purchases live in a stack of paper bills. Cash flow lives in the shopkeeper's head.
This isn't a failure of the business. It's a failure of the tools available to it. Bookkeeping software was built for people who already think in spreadsheets — and that assumption quietly excludes millions of businesses that are otherwise completely viable. PwC estimates AI could unlock USD 135–150 billion in value for manufacturing MSMEs alone by 2035, and a meaningful share of that value is locked behind a simple, unglamorous problem: these businesses don't have a practical way to turn their daily activity into structured data.
That's the gap we set out to close with STACKR — an AI-powered, multimodal accounting and financial record-building platform for informal MSMEs, built entirely on AWS.
The Core Idea: Don't Make the Shopkeeper Adapt to the Software
Most fintech tools for small businesses fail for the same reason: they ask the business owner to change how they work. Learn a new app. Enter data manually. Understand line items and categories.
STACKR flips that. Instead of asking a shop owner to learn accounting, it meets them where they already are:
Photograph a bill — a paper receipt from a supplier, and it becomes a structured transaction.
Say it out loud — "I bought 2 kg tomato from Ramu for Rs 200," and it becomes a structured transaction.
No forms. No categories to select. No new habits to build. The AI does the translation work; the shopkeeper just keeps doing what they already do.
How AWS Made This Possible
This was built for an AWS hackathon under the Ship It track, and one of the most valuable parts of the process was discovering how well AWS's managed AI services compose into a single, coherent pipeline — without us having to build any of the underlying models ourselves.
Amazon S3 is the front door. Bill photos and voice recordings, captured on the user's phone, go straight into S3 via presigned URLs — no large media files ever touch our compute layer directly.
Amazon Textract, specifically the AnalyzeExpense API, does the heavy lifting on the document side. Unlike generic OCR, AnalyzeExpense already understands the shape of a receipt — it returns vendor names, line items, quantities, and totals in a semi-structured form. That one API choice removed an enormous amount of parsing work we would otherwise have had to write ourselves.
Amazon Transcribe handles the voice path, converting a spoken sentence into text so a shopkeeper can log a sale or purchase without typing a single character.
Amazon Bedrock, running Claude, is where both paths converge. Whether the input started as a Textract-parsed receipt or a Transcribe transcript, it flows through a single shared prompt that normalizes it into one consistent transaction schema: item, quantity, price, counterparty, date. Two very different input modalities, one data model — that unification is what let a three-person team ship something coherent in a hackathon timeframe instead of building two parallel systems.
Amazon DynamoDB stores the resulting transactions with minimal operational overhead — exactly right for the transaction volume of a single small business.
AWS Lambda and Amazon API Gateway form the entire compute and API surface: one function per pipeline step, all serverless, all independently deployable. This let our backend and frontend work happen in parallel almost from day one.
AWS Amplify hosts and deploys the frontend with continuous deployment straight from our repo — which mattered enormously for a Ship It submission that had to end with a live, working URL, not just a local demo.
What It Actually Produces
Every transaction — whether it came from a photo or a voice note — lands in the same structured record. Over time, that accumulation becomes something these businesses have rarely had: a real financial history. STACKR turns that history into:
A live profit-and-loss view, updated as transactions come in
Exportable CSV and PDF reports formatted for external use — the kind of document a lender can actually work with
A Record Readiness indicator, which shows how complete a business's documentation currently is for a formal loan application
That last piece was a deliberate scoping decision, and it's one we'd make again. Record Readiness tells a business owner what they have and what's missing — it does not score, rank, or predict creditworthiness. The application's job is to make a business's activity visible and verifiable. The judgment of whether to lend stays exactly where it belongs: with the lender.
What We Learned
Building an actual end-to-end AI pipeline on AWS, rather than wiring up a single API call, changed how we think about the AWS AI stack. A few things stood out:
Purpose-built beats general-purpose, where it's available. Textract's AnalyzeExpense did more of our parsing work than a generic OCR call plus an LLM prompt would have, because it already understands receipts as a document type.
A shared foundation model is a powerful unifier. Routing two very different inputs (image-derived text and speech-derived text) through one Bedrock prompt into one schema kept our data model — and our codebase — from splitting into two systems.
Serverless removes an entire category of hackathon risk. No servers to provision meant no servers to misconfigure under time pressure, and Lambda's function-per-step model meant our backend and frontend teammates could build against a stable API contract without waiting on each other.
Where It Goes From Here
STACKR is a working proof of concept, not a finished product — the Record Readiness logic, multilingual voice support, and broader receipt-format coverage are all areas we'd deepen next. But the core thesis held up: informal businesses don't need to be taught accounting. They need their existing activity translated into a form the formal financial system can read. AWS gave us the pieces to build that translation layer in a weekend, and to actually ship it.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article